Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Low-resource keyword spotting using contrastively trained transformer acoustic word embeddings

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A transformer encoder trained with a contrastive loss produces acoustic word embeddings that outperform all compared approaches for very low-resource keyword spotting in Luganda and Bambara.

desk verdict A transparent, incrementally useful KWS paper: contrastive transformer AWEs show consistent but modest gains, with test-set evidence incomplete for the headline claim. read the letter →

arxiv 2506.17690 v1 pith:4XGPFRNH submitted 2025-06-21 eess.AS

Julian Herreilers , Christiaan Jacobs , Thomas Niesler This is my paper · ORCID
classification eess.AS
keywords Keywordspottingacousticwordembeddingscontrastivelearningtransformerencoderlow-resourcespeechapplicationsquery-by-exampleself-supervisedrepresentationsLuganda
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to show that query-by-example keyword spotting in very low-resource languages can be made more accurate by using a transformer encoder trained directly in the embedding space with a contrastive loss. The proposed ContrastiveTransformer requires no transcribed audio in the target language: it is trained on well-resourced languages and then applied to Luganda and Bambara radio broadcasts using only a small collection of isolated keyword templates as queries. The paper reports that on the test sets the model outperforms every considered baseline, including a recurrent contrastive model, a reconstruction-based autoencoder, direct pooling of self-supervised features, and a dynamic time warping system. These results matter because humanitarian monitoring systems often must be deployed quickly in languages with no ASR and almost no labelled data.

What carries the argument

The load-bearing object is the ContrastiveTransformer, an encoder-only transformer whose acoustic word embedding is taken from the first output position after a prepended trainable vector. Training is driven by the NT-Xent contrastive loss: within each batch, embeddings of two different utterances of the same word type are pulled together under cosine similarity while embeddings of different word types are pushed apart, with a temperature parameter $\tau$ controlling the sharpness. Inputs are frame-level features from layer 10 of mHuBERT-147, a compact multilingual self-supervised model; the paper finds this layer matches XLS-R layer 13 for unseen languages while being better for a seen language. The machinery therefore replaces the reconstruction objective of older AWE models with a direct geometric objective, producing an embedding space tailored to the cosine-distance comparisons used in keyword spotting.

What would settle it

Give the ContrastiveRNN and CAE-RNN the same development-set hyperparameter search and input-feature layer selection as the ContrastiveTransformer; if either baseline then matches or exceeds the transformer's MAP on the Luganda and Bambara test sets, the reported advantage would be tuning rather than architecture.

Watch

Extended reading notes

Core claim

The paper's central claim is that directly optimising a transformer encoder with the normalised temperature-scaled cross entropy (NT-Xent) loss produces acoustic word embeddings that transfer across languages and improve very low-resource keyword spotting. A trainable vector is prepended to each input feature sequence, and the first output vector of the final transformer layer, projected to 256 dimensions, becomes the embedding. Trained on 400,000 word pairs from four well-resourced languages, the model encodes keyword templates and sliding windows from Luganda and Bambara radio speech without any target-language adaptation. On the test set the ContrastiveTransformer reaches a mean average precision of 65.3% versus 60.6% for the ContrastiveRNN in Luganda, and 69.9% versus 68.9% in Bambara, improving on every reported metric.

Load-bearing premise

The comparison to the baselines is fair even though the ContrastiveTransformer was tuned on the development set and the contrastive RNN and autoencoder baselines were not given equivalent tuning.

Editorial extensions

If this is right

  • A new low-resource language could get a keyword spotting system with only a small set of isolated keyword templates and no transcribed target-language audio.
  • The transfer appears to work for languages not present in the pre-trained feature model's training data, since Bambara is not in mHuBERT-147 or XLS-R, and the model still improves over baselines.
  • The consistent improvement in P@N means the transformer embedding recovers more of the true keyword occurrences, not just the top-ranked matches.
  • When choosing AWE training languages, a related language helps when available; when none is available, diverse languages are better than several closely related ones.
  • Contrastive training on top of large self-supervised features yields a consistent gain over using those features directly by meanpooling or subsampling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because only the ContrastiveTransformer's architecture and input-feature layer were selected on the development sets, while the baselines kept their original configurations, the reported margin may partly reflect tuning effort; an equal-tuning comparison would clarify the architectural advantage.
  • The embeddings produced are generic fixed-dimensional representations, so the same contrastive transformer could be tested on other speech tasks such as hate-speech detection or spoken term discovery, which would show whether the benefit is task-specific.
  • Testing on additional language families outside Bantu and Mande would indicate how much of the success depends on mHuBERT-147 having seen related languages during its self-supervised pretraining.
  • A systematic variation of the number of keyword templates would quantify the labelled effort a humanitarian deployer actually needs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces the ContrastiveTransformer, an encoder-only transformer that is trained with the NT-Xent contrastive loss to produce acoustic word embeddings (AWEs) for query-by-example keyword spotting in very low-resource languages. The authors train on well-resourced and related languages (using 400k word pairs from NCHLT and Swahili) and evaluate on radio broadcast search corpora in English, Luganda, and Bambara. They compare against a correspondence autoencoder RNN, meanpooling and subsampling of self-supervised features, a contrastive RNN, and a DTW baseline. The main claims are that the ContrastiveTransformer outperforms all considered approaches on the development set and outperforms the two approaches reported on the test set.

Significance. If the result holds, the paper makes a practically useful contribution: a simple transformer-based AWE model that needs only a few isolated keyword templates and no transcribed target-language data, and that improves keyword spotting on under-resourced African languages. The paper is also one of the few to compare several AWE approaches directly on radio-broadcast keyword spotting. The analysis of training-language combinations in Section 6.3 is a useful practical result. The main limitations are statistical: the test-set evidence is incomplete and no significance testing is reported, which weakens the strength of the headline claim.

major comments (4)
  1. [Section 6.2, Table 4] The abstract and conclusion claim that the proposed approach offers performance improvements over all considered existing approaches, but Table 4 reports test-set results only for ContrastiveRNN and DTW. CAE-RNN, meanpooling, and subsampling are absent from the test set, so the 'all considered' claim is not directly tested. Please either report test-set MAP, P@10, and P@N for all five baselines, or restrict the claim to the systems actually compared on the test set.
  2. [Section 6.2, Table 4] No significance tests, confidence intervals, or multiple-seed variation are reported. The Bambara MAP gap between ContrastiveTransformer and ContrastiveRNN is 1.0 percentage point (69.9 vs 68.9) and the P@10 gap is 0.9 percentage point (82.2 vs 81.3); these differences are small relative to expected run-to-run variability in neural model training. Please provide per-keyword paired tests (for example, a paired bootstrap or Wilcoxon signed-rank test over keyword types) or report variance across training seeds, so the reader can judge whether the observed test-set superiority is statistically distinguishable from noise.
  3. [Section 5.2] The paper states that the ContrastiveTransformer architecture (3 layers, 16 heads, 256-dimensional embedding) 'was determined by optimisation on our development sets' and that the mHuBERT layer-10 input features were chosen in preliminary experiments, while the baselines are used as originally proposed without analogous target-language tuning. If the development sets used for this selection include Luganda and Bambara, the comparison is biased in favor of the proposed model. Please clarify exactly which development sets were used for architecture and feature-layer selection, and consider reporting results with a fixed a-priori architecture or a sensitivity analysis to show how much the development-set advantage depends on this tuning.
  4. [Section 5.2 vs Table 3] Section 5.2 says that using the outputs of the 10th transformer layer of mHuBERT achieves good performance, yet Table 3 reports meanpooling and subsampling results at layer 8 and Figure 2 indicates layer 8 as the best meanpooling layer. It is therefore unclear whether the ContrastiveTransformer and the other AWE models use layer 8 or layer 10 as input features. Please state explicitly which mHuBERT layer is used as the input to each of the trained AWE models and whether this is the same layer for all models, so the comparison is not confounded by a difference in input representation.
minor comments (4)
  1. [Section 6.2] The sentence 'In contrast to the the two under-resourced languages' contains a duplicated 'the'.
  2. [Tables 1 and 2] Duration entries in Tables 1 and 2 have inconsistent spacing (for example, '34 .27m' and '2 .05h'); please format them consistently, for example as '34.27m' and '2.05h'.
  3. [Section 6.3, Table 5] Table 5 reports models trained with 100K pairs per language while Table 3 uses 400K pairs total; please state explicitly whether the model architecture and input features are identical in both tables so the comparison across training-language combinations is unambiguous.
  4. [Figure 2] It would be helpful to mark the layer selected for input features to the ContrastiveTransformer (and to the other trained AWE models) in Figure 2, so the reader can directly see how that layer compares with the meanpooling results.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning: the paper is an empirical benchmark with held-out test-set evaluation; development-set tuning is model selection, not a fitted prediction.

full rationale

I examined the claimed derivation chain and found no step in which a prediction or first-principles result is equivalent to its inputs by construction. The paper's central claim is an empirical comparison: the ContrastiveTransformer is trained on NCHLT word pairs and evaluated on held-out test sets for Luganda and Bambara (Table 4). Section 5.2 states that the ContrastiveTransformer architecture was 'determined by optimisation on our development sets', while Table 4 reports results on test sets not used for that optimisation. This is ordinary model selection, not a fitted parameter renamed as a prediction. The self-citations, such as [14] for ContrastiveRNN and [17] for training on related languages, are used as baselines or experimental design choices; the central superiority claim is not derived from those citations alone and is checked against an independent test set. The absence of significance tests or variance estimates is a statistical reporting concern, but it is not circularity. No equation in the paper defines the proposed model's output in terms of the evaluation metric, and no baseline result is folded into the training objective. Therefore the derivation is self-contained and the paper receives a circularity score of 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the choice of transformer architecture and feature layer via development set tuning, on the validity of cross-speaker same-word pairs, and on the transferability of mHuBERT features. No new entities are introduced. The main free parameters are the model configuration, the selected feature layer, the keyword frequency filter, the unreported NT-Xent temperature, and the search window settings.

free parameters (5)
  • Transformer encoder architecture = 3 layers, 16 attention heads, 256-dim AWE
    Determined by optimization on development sets (Section 5.2); performance of the proposed model depends on this choice, and the baselines were not given the same tuning.
  • mHuBERT input feature layer = layer 10
    Selected in preliminary experiments as achieving the same unseen-language performance as XLS-R layer 13 and better performance for seen languages (Section 5.2).
  • Keyword frequency filter = keywords with at least 10 occurrences; 12 Luganda, 23 Bambara keywords
    Adopted to make P@10 meaningful (Section 5.3); this filters out rare keywords and affects the averaged scores.
  • NT-Xent temperature tau = not reported
    Temperature in the contrastive loss (Eq. 1) controls the sharpness of the distribution; its value is not stated, so its contribution to the results cannot be assessed.
  • Search window sweep settings = length 10-65 frames, stride 5 frames
    The search window is swept over these settings (Section 5.3), but the way overlapping segment scores are combined for the final similarity is not fully specified, making this an unstated control setting.
assumptions (4)
  • domain assumption Same-word pairs across speakers are valid positive pairs for contrastive training.
    The training procedure (Section 5.2) assumes that two different audio instances of the same word type should map to the same embedding, which is the standard AWE assumption.
  • domain assumption The development set is representative of the test set for architecture selection.
    The ContrastiveTransformer's architecture and feature layer are chosen on development sets (Section 5.2), and the test set is then used for final evaluation; if the dev and test distributions differ, the selected configuration may not generalize.
  • domain assumption Montreal Forced Aligner alignments for Swahili are sufficiently accurate to create training pairs.
    Word-level alignments for Swahili were obtained using MFA (Section 5.2); errors in alignments could degrade the quality of positive pairs.
  • domain assumption mHuBERT features are effective for languages not in its training data (Bambara).
    The system relies on mHuBERT-147 features (Section 5.1); the paper states Bambara is not in the training set, so transferability is assumed rather than proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Low-resource keyword spotting using contrastively trained transformer acoustic word embeddings." pith.science (2026). https://pith.science/paper/4XGPFRNH

@misc{pith2026250617690,
  author       = {Pith},
  title        = {Pith review of: Low-resource keyword spotting using contrastively trained transformer acoustic word embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XGPFRNH}},
  note         = {Machine review of arXiv:2506.17690}
}
read the original abstract

We introduce a new approach, the ContrastiveTransformer, that produces acoustic word embeddings (AWEs) for the purpose of very low-resource keyword spotting. The ContrastiveTransformer, an encoder-only model, directly optimises the embedding space using normalised temperature-scaled cross entropy (NT-Xent) loss. We use this model to perform keyword spotting for radio broadcasts in Luganda and Bambara, the latter a severely under-resourced language. We compare our model to various existing AWE approaches, including those constructed from large pre-trained self-supervised models, a recurrent encoder which previously used the NT-Xent loss, and a DTW baseline. We demonstrate that the proposed contrastive transformer approach offers performance improvements over all considered existing approaches to very low-resource keyword spotting in both languages.

Figures

Figures reproduced from arXiv: 2506.17690 by the authors.

Figure 1
Figure 1. ContrastiveTransformer architecture. A trainable vector is prepended to the input sequence and processed by the transformer. The first output vector of the final layer is used as the AWE w. The model is trained using a batch of distinct pairs, where each pair consists of two different examples of the same word type. During training, the distance between two em￾beddings of the same word type is minimised while the di… view at source ↗
Figure 2
Figure 2. KWS results on the development set. AWEs are obtained through meanpooling features extracted from different transformer layers. The baseline corresponds to DTW using bottleneck features [7] [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 2
Figure 2. The graphs show that choosing the correct layer at which [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 38 canonical work pages

  1. [1]

    Introduction The United Nations have for a number of years funded the de- velopment of keyword spotting (KWS) systems that can be ap- plied to radio broadcasts as a way for the organisation to moni- tor and inform its humanitarian relief efforts [1, 2]. These sys- tems operate in areas of the developing world with little or no in- ternet infrastructure, w...

    work page

  2. [2]

    One of these five is the Con- trastiveTransformer we propose

    Acoustic word embedding (A WE) techniques We evaluate and compare five AWE modelling approaches for QbE using radio speech. One of these five is the Con- trastiveTransformer we propose. The other four are (i) a corre- spondence autoencoder RNN (CAE-RNN) as proposed in [11] and applied to hate speech detection in [20], (ii) meanpooling and (iii) subsamplin...

    arXiv 2025

  3. [3]

    Keyword spotting using A WEs When performing keyword spotting (KWS) by QbE, a collec- tion of isolated keyword templates is used to query whether the associated keyword is present in an unlabelled utterance [20]. Since word boundaries are not available for the search utterance, a variable-length window can be applied to divide the utterance into overlappi...

    work page

  4. [4]

    Datasets Different datasets are used for AWE training and for subsequent KWS. 4.1. A WE training For the supervised training of multilingual AWE models, as re- quired for methods 2.1, 2.3 and 2.4 in Section 2, we utilise the NCHLT dataset [23]. This corpus contains the 11 official spoken languages of South Africa, 9 of which belong to the Southern Bantu f...

    work page

  5. [5]

    Pre-trained models For meanpooling and subsampling as described in Section 2.2, we consider four pre-trained self-supervised models from the wav2vec2.0 and HuBERT families [25, 26]

    Experimental setup 5.1. Pre-trained models For meanpooling and subsampling as described in Section 2.2, we consider four pre-trained self-supervised models from the wav2vec2.0 and HuBERT families [25, 26]. For the former, 2 4 6 8 10 12 14 16 18 20 22 24 Transformer layer 0 20 40 60 80MAP (%) English 2 4 6 8 10 12 14 16 18 20 22 24 Transformer layer Lugand...

    work page 1920

  6. [6]

    Results 6.1. Keyword spotting by meanpooling of self-supervised features A layer-wise analysis of self-supervised model features, similar to that performed in [18] for word discrimination, is shown in Fig. 2. The graphs show that choosing the correct layer at which feature meanpooling is applied leads to improvements over the DTW baseline for all three la...

    work page

  7. [7]

    The AWEs produced by this transformer are then used to encode speech in the target language

    Conclusion We propose a new approach to very low-resource keyword spotting (KWS), the ContrastiveTransformer, which leverages acoustic representations computed by large pre-trained archi- tectures and performs subsequent contrastive training of acous- tic word embeddings (AWEs) using well-resourced languages. The AWEs produced by this transformer are then...

    work page

  8. [8]

    We also thank the DW Ackermann Bursary Fund and Telkom South Africa for support

    Acknowledgements We thank NVIDIA for donating GPU equipment, Ewald van der Westhuizen for the NCHLT alignments and Herman Kamper for valuable discussions. We also thank the DW Ackermann Bursary Fund and Telkom South Africa for support

    work page

Show all 39 references
  1. [9]

    UN GLOBAL PULSE - UNHCR Innovation Service, Social Me- dia and Forced Displacement: Big Data Analytics & Machine- Learning - White Paper, 2017

    2017

  2. [10]

    Feature learning for efficient ASR-free keyword spot- ting in low-resource languages,

    E. van der Westhuizen, H. Kamper, R. Menon, J. Quinn, and T. Niesler, “Feature learning for efficient ASR-free keyword spot- ting in low-resource languages,”Computer Speech and Language, vol. 71, p. 101275, 2022

    2022

  3. [11]

    Unsupervised spoken keyword spotting via segmental DTW on Gaussian posteriorgrams,

    Y . Zhang and J. R. Glass, “Unsupervised spoken keyword spotting via segmental DTW on Gaussian posteriorgrams,” inProceedings of IEEE Workshop on Automatic Speech Recognition and Under- standing, Moreno, Italy, 2009

    2009

  4. [12]

    Fast ASR-free and almost zero-resource keyword spotting using DTW and CNNs for humanitarian monitoring,

    R. Menon, H. Kamper, J. Quinn, and T. Niesler, “Fast ASR-free and almost zero-resource keyword spotting using DTW and CNNs for humanitarian monitoring,” in Proceedings of Interspeech, Hy- derabad, India, 2018

    2018

  5. [13]

    ASR- free CNN-DTW keyword spotting using multilingual bottleneck features for almost zero-resource languages,

    R. Menon, H. Kamper, E. Yilmaz, J. Quinn, and T. Niesler, “ASR- free CNN-DTW keyword spotting using multilingual bottleneck features for almost zero-resource languages,” in Proceedings of Workshop on Spoken Language Technologies for Under-resourced Languages (SLTU), Gurugram, ...

    2018

  6. [14]

    Feature exploration for almost zero-resource asr-free keyword spotting using a multilingual bottleneck extractor and correspondence autoencoders,

    R. Menon, H. Kamper, E. van der Westhuizen, J. Quinn, and T. Niesler, “Feature exploration for almost zero-resource asr-free keyword spotting using a multilingual bottleneck extractor and correspondence autoencoders,” in Proceedings of Interspeech , Graz, Austria, 2019

    2019

  7. [15]

    Low-resource ASR-free key- word spotting using listen-and-confirm,

    E. van der Westhuizen, M. Ribeiro, J. Jansen van V ¨uren, P. Hidalgo-Sanchis, and T. Niesler, “Low-resource ASR-free key- word spotting using listen-and-confirm,” inProceedings of the An- nual Meeting of the Special Interest Group on Under-resourced Languages (SIGUL), Dublin, ...

    2023

  8. [16]

    Fixed- dimensional acoustic embeddings of variable-length segments in low-resource settings,

    K. D. Levin, K. Henry, A. Jansen, and K. Livescu, “Fixed- dimensional acoustic embeddings of variable-length segments in low-resource settings,” Proceedings of the Workshop on Auto- matic Speech Recognition and Understanding (ASRU), 2013

    2013

  9. [17]

    Acoustic span embeddings for multilingual query-by-example search,

    Y . Hu, S. Settle, and K. Livescu, “Acoustic span embeddings for multilingual query-by-example search,” in Proceedings of the IEEE Spoken Language Technology Workshop (SLT), Virtual, 2021, pp. 935–942

    2021

  10. [18]

    Audio word2vec: Unsupervised learning of audio segment repre- sentations using sequence-to-sequence autoencoder,

    Y .-A. Chung, C.-C. Wu, C.-H. Shen, H. yi Lee, and L.-S. Lee, “Audio word2vec: Unsupervised learning of audio segment repre- sentations using sequence-to-sequence autoencoder,” in Proceed- ings of Interspeech, San Francisco, USA, 2016

    2016

  11. [19]

    Truly unsupervised acoustic word embeddings using weak top-down constraints in encoder-decoder models,

    H. Kamper, “Truly unsupervised acoustic word embeddings using weak top-down constraints in encoder-decoder models,” in Pro- ceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP) , Brighton, United Kingdm, 2018, pp. 6535–3539

    2018

  12. [20]

    Discriminative acoustic word embed- dings: recurrent neural network-based approaches,

    S. Settle and K. Livescu, “Discriminative acoustic word embed- dings: recurrent neural network-based approaches,” in Proceed- ings of the IEEE Spoken Language Technology Workshop (SLT) , San Juan, Peurto Rico, 2016

    2016

  13. [21]

    A comparison of self-supervised speech representations as input features for unsupervised acoustic word embeddings,

    L. van Staden and H. Kamper, “A comparison of self-supervised speech representations as input features for unsupervised acoustic word embeddings,” Proceedings of the IEEE Spoken Language Technology Workshop (SLT), 2021

    2021

  14. [22]

    Acoustic word em- beddings for zero-resource languages using self-supervised con- trastive learning and multilingual adaptation,

    C. Jacobs, Y . Matusevych, and H. Kamper, “Acoustic word em- beddings for zero-resource languages using self-supervised con- trastive learning and multilingual adaptation,” in Proceedings of the Spoken Language Technology (SLT) Workshop , Shenzhen, China, 2021

    2021

  15. [23]

    Self-supervised acoustic word embedding learning via correspondence transformer encoder,

    J.-H. Lin, X. Yue, J. Ao, and H. Li, “Self-supervised acoustic word embedding learning via correspondence transformer encoder,” in Proceedings of Interspeech, Dublin, Ireland, 2023

    2023

  16. [24]

    Improved acous- tic word embeddings for zero-resource languages using multilin- gual transfer,

    H. Kamper, Y . Matusevych, and S. Goldwater, “Improved acous- tic word embeddings for zero-resource languages using multilin- gual transfer,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 1107–1118, 2021

    2021

  17. [25]

    Multilingual transfer of acoustic word embeddings improves when training on languages related to the target zero-resource language,

    C. Jacobs and H. Kamper, “Multilingual transfer of acoustic word embeddings improves when training on languages related to the target zero-resource language,” in Proceedings of Interspeech , Brno, Czech Republic, 2021

    2021

  18. [26]

    Analyzing acoustic word embeddings from pre-trained self-supervised speech models,

    R. Sanabria, H. Tang, and S. Goldwater, “Analyzing acoustic word embeddings from pre-trained self-supervised speech models,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Rhodes Island, Greece, 2023

    2023

  19. [27]

    Acoustic word embeddings for untranscribed target languages with con- tinued pretraining and learned pooling,

    R. Sanabria, O. Klejch, H. Tang, and S. Goldwater, “Acoustic word embeddings for untranscribed target languages with con- tinued pretraining and learned pooling,” in Proceedings of Inter- speech, Dublin, Ireland, 2023

    2023

  20. [28]

    Towards hate speech detection in low-resource languages: Comparing ASR to acoustic word embeddings on Wolof and Swahili,

    C. Jacobs, N. C. Rakotonirina, E. A. Chimoto, B. A. Bassett, and H. Kamper, “Towards hate speech detection in low-resource languages: Comparing ASR to acoustic word embeddings on Wolof and Swahili,” in Proceedings of Interspeech, Dublin, Ire- land, 2023

    2023

  21. [29]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proceedings of the International Conference on Neural Infor- mation Processing Systems (NIPS), 2017

    2017

  22. [30]

    BERT: pre- training of deep bidirectional transformers for language under- standing,

    J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre- training of deep bidirectional transformers for language under- standing,” in Proceedings of the Conference of the North Amer- ican Chapter of the Association for Computational Linguis- tics: Human Language Technologie...

    2019

  23. [31]

    The NCHLT speech corpus of the South African lan- guages,

    E. Barnard, M. Davel, C. van Heerden, F. de Wet, and J. Baden- horst, “The NCHLT speech corpus of the South African lan- guages,” in Proceedings of the Workshop Spoken Language Tech- nologies for Under-resourced Languages (SLTU), St Petersburg, Russia, 2014

    2014

  24. [32]

    Common Voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common Voice: A massively-multilingual speech corpus,” in Proceedings of the International Conference on Language Re- sources and Evaluation (LREC), Marseille...

    2020

  25. [33]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” in Proceedings of NeurIPS, 2020

    2020

  26. [34]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., 2021

    2021

  27. [35]

    XLS-R: Self-supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Con- neau, and M. Auli, “XLS-R: Self-supervised cross-lingual speech representation learning at scale,” in Proceedings of Interspeech , Incheon, Korea, 2022

    2022

  28. [36]

    mhubert-147: A compact multilingual hubert model,

    M. Z. Boito, V . Iyer, N. Lagos, L. Besacier, and I. Calapodescu, “mhubert-147: A compact multilingual hubert model,” in Pro- ceedings of Interspeech, Kos Island, Greece, 2024

    2024

  29. [37]

    Montreal forced aligner: Trainable text-speech align- ment using kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal forced aligner: Trainable text-speech align- ment using kaldi,” in Interspeech, Stockholm, Sweden, 2017

    2017

  30. [38]

    Rapid evaluation of speech representations for spoken term discovery,

    M. A. Carlin, S. Thomas, A. Jansen, and H. Hermansky, “Rapid evaluation of speech representations for spoken term discovery,” in Proceedings of Interspeech, Florence, Italy, 2011

    2011

  31. [39]

    Layer-wise analysis of a self-supervised speech representation model,

    A. Pasad, J.-C. Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” Proceedings of the IEEE Automatic Speech Recognition and Understanding Work- shop (ASRU), 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.